Cloud Computing Economies of Scale AWS Executive Symposium 2010 James Hamilton, 2010/7/15 VP & Distinguished Engineer, Amazon Web Services email: James@amazon.com web: mvdirona.com/jrh/work blog: perspectives.mvdirona.com
Agenda Follow the money in infrastructure Infrastructure cost breakdown Power Distribution Efficiency Mechanical System Efficiency Server Design & Utilization Sea Change in Networking Cloud Computing Economics Why utility computing makes sense economically 2010/7/15 http://perspectives.mvdirona.com 2
Economies of Scale 2006 comparison of very large service with mid-size: (~1000 servers): Large Service [$13/Mb/s/mth]: $0.04/GB Medium [$95/Mb/s/mth]: $0.30/GB (7.1x) Large Service: $4.6/GB/year (2x in 2 Datacenters) Medium: $26.00/GB/year* (5.7x) Large Service: Over 1.000 servers/admin Enterprise: ~140 servers/admin (7.1x) Large block h/w purchases significantly more economic Large weekly purchases offer significant savings H/W Manufacturers willing & able to do custom designs at scale Automation & custom s/w investments amortize well at scale Summary: scale economics strongly in play 2010/7/15 http://perspectives.mvdirona.com 3
Where Does the Money Go? Assumptions: Facility: ~$88M for 8MW facility Servers: Roughly 46k @ $1.45k each Server power draw at 30% load: 80% Commercial Power: ~$0.07/kWhr PUE: 1.5 21% 13% Monthly Costs 5% Servers 54% Networking Equipment Power Distribution & Cooling Power 8% Other Infrastructure 3yr server, 4yr net gear, & 10 yr infrastructure amortization Observations: 34% costs functionally related to power (trending up while server costs down) Networking high at 8% of costs & 19% of total server cost 2010/7/15 Updated from: http://perspectives.mvdirona.com/2008/11/28/costofpowerinlargescaledatacenters.aspx http://perspectives.mvdirona.com 4
Agenda Follow the money in infrastructure Infrastructure cost breakdown Power Distribution Efficiency Mechanical System Efficiency Server Design & Utilization Sea Change in Networking Cloud Computing Economics Why utility computing makes sense economically 2010/7/15 http://perspectives.mvdirona.com 5
208V Power Distribution High Voltage Utility Distribution 11% loss in distribution.997*.94*.98*.98*.99 = 89% 2.5MW Generator (180 gal/hr) IT Load (servers, storage, Net, ) 115kv Sub-station 13.2kv UPS: Rotary or Battery UPS & Gen often on 480v Transformers Transformers ~1% loss in switch gear & conductors 13.2kv 13.2kv 480V 0.3% loss 6% loss 2% loss 2% loss 99.7% efficient 94% efficient, ~97% available 98% efficient 98% efficient 2010/7/15 http://perspectives.mvdirona.com 6
Cold Hot Conventional Mechanical Design Blow down & Evaporative Loss at 8MW facility: ~200,000 gal/day Cooling Tower CWS Pump Heat Exchanger (Water-Side Economizer) A/C Condenser A/C Compressor Primary Pump A/C Evaporator Server fans 6 to 9W each Diluted Hot/Cold Mix leakage cold fans Overall Mechanical Losses ~22% Computer Room Air Handler Air Impeller Air-side Economization 2010/7/15 http://perspectives.mvdirona.com 7
Hot Aisle/Cold Aisle Containment Intel WriteLine Intel 2010/7/15 http://perspectives.mvdirona.com 8
Air-Side Economization & Evaporative Cooling Limiting factors to high temp operation Higher fan power trade-off More semiconductor leakage current Possible negative failure rate impact Avoid direct expansion cooling entirely Air side economization Higher data center temperatures Evaporative cooling Requires Filtration Particulate & chemical pollution 2010/7/15 http://perspectives.mvdirona.com 9
Server Innovation Shared Infrastructure Racks Shared redundant PSUs & fans e.g. Dell Fortuna & Rackable CloudRack Next Level: Multi-server on board Intel Atom: SeaMicro ARM: SmoothStone Very Low-Cost, Low-Power Servers ARM, Atom, client & embedded CPUs Cold storage (reduce CPU $ to GB) Highly partitionable workloads: Web services, memcached Low utilization is still the elephant in room 2010/7/15 http://perspectives.mvdirona.com 10
Sea Change in Net Gear Current networks over-subscribed Forces workload placement restrictions Goal: all points in datacenter equidistant Mainframe model goes commodity Competition at each layer rather than vertical integration OpenFlow: open S/W platform Distributed control plane to central control E.g. VL2, Portland, and others 2010/6/23 http://perspectives.mvdirona.com 11
Agenda Follow the money in infrastructure Infrastructure cost breakdown Power Distribution Efficiency Mechanical System Efficiency Server Design & Utilization Sea Change in Networking Cloud Computing Economics Why utility computing makes sense economically 2010/7/15 http://perspectives.mvdirona.com 12
Infrastructure at Scale Datacenter design efficiency Average datacenter efficiency low with PUE over 2.0 (Source: EPA) Many with PUE well over 3.0 High scale cloud services in the 1.2 to 1.5 range Lowers computing cost & better for environment Multiple datacenters At scale multiple datacenters can be used Close to customer Cross datacenter data redundancy Address international markets efficiently Avoid massive upfront data cost & years to fully utilize 2010/7/15 http://perspectives.mvdirona.com 13
Broad set of services: Infrastructure Services SimpleDB AWS Approach Simple Storage Service CloudFront Simple Queue Service Elastic MapReduce Relational Database Service Elastic Block Store Premium Support Virtual Private Cloud Payments & Billing Flexible Payment Services DevPay On Demand Workforce Mechanical Turk Alexa Web Services Web Information Service Top Sites Merchant Services Fulfillment Web Service Open the hood approach Simple, layerable building block services Component services are substitutable 11/10/2009 http://perspectives.mvdirona.com 14
Amazon Cycle of Innovation 15+ years of operational excellence Managing secure, highly available, multi-datacenter infrastructure Experienced at low margin cycle of innovation: Innovate Listen to customers Drive down costs & improve processes Pass on value to customers AWS price reductions expected to continue 11/10/2009 http://perspectives.mvdirona.com 15
AWS Scale Looking at Quantcast Top 500k sites (11/2009) 53% of cloud hosted sites are using AWS 27% growth in AWS hosted site count Oct to Nov More sites than all others combined High growth workloads: data Intensive computing, commercial HPC, analysis, & optimization Very high scale supports deep investment in automation, monitoring, operations, & funds faster innovation Sources: http://www.quantcast.com/top-sites-1 http://www.jackofallclouds.com/2009/11/state-of-the-cloud-november-2009/ 11/10/2009 http://perspectives.mvdirona.com 16
Utilization & Economics Server utilization problem 30% utilization VERY good &10% to 20% common Expensive & not good for environment Solution: pool number of heterogeneous services Single reserve capacity pool far more efficient Non-correlated peaks & law of large numbers Pay as you go & pay as you grow model Don t block the business Don t over buy Transfers capital expense to variable expense Apply capital for business investments rather than infrastructure Charge back models drive good application owner behavior Cost encourages prioritization of work by application developers High scale needed to make a market for low priority work 2010/7/15 http://perspectives.mvdirona.com 17
Amazon Web Services Pace of Innovation» Reserved Instances in EU Region» Elastic MapReduce» SQS in EU Region» New SimpleDB Features» FPS General Availability» AWS Security Center» AWS Multi-Factor Authentication» Virtual Private Cloud» Lower Reserved Instance Pricing» Relational Database Service» High-Memory Instances» Lower EC2 Pricing» EC2 with Windows Server 2008, Spot Instances, Boot from Amazon EBS» CloudFront Streaming» VPC enters Unlimited Beta» AWS Region in Northern California» AWS Import/Expert International Support» Amazon EC2 with Windows» Amazon EC2 in EU Region» AWS Toolkit for Eclipse» Amazon EC2 Reserved Instances» AWS Import/Export» Lower CloudFront pricing tiers» AWS Management Console» New CloudFront Feature» Monitoring, Auto Scaling & Elastic Load Balancing» Elastic MapReduce in EU» EBS Shared Snapshots» SimpleDB in EU Region» Monitoring, Auto Scaling & Elastic Load Balancing in EU» CloudFront private content» SAS70 Type II Audit» AWS SDK for.net» EC2 Reserved Instances with Windows, Extra Large High Memory Instances» S3 Versioning» AWS Consolidated Billing» Lower pricing for Outbound Data 2010/7/15 http://perspectives.mvdirona.com 18
More Information This Slide Deck: I will post slides to http://mvdirona.com/jrh/work later this week Berkeley Above the Clouds http://perspectives.mvdirona.com/2009/02/13/berkeleyabovetheclouds.aspx Degraded Operations Mode http://perspectives.mvdirona.com/2008/08/31/degradedoperationsmode.aspx Cost of Power http://perspectives.mvdirona.com/2008/11/28/costofpowerinlargescaledatacenters.aspx http://perspectives.mvdirona.com/2008/12/06/annualfullyburdenedcostofpower.aspx Power Optimization: http://labs.google.com/papers/power_provisioning.pdf Cooperative, Expendable, Microslice Servers http://perspectives.mvdirona.com/2009/01/15/thecaseforlowcostlowpowerservers.aspx Power Proportionality http://www.barroso.org/publications/ieee_computer07.pdf Resource Consumption Shaping: http://perspectives.mvdirona.com/2008/12/17/resourceconsumptionshaping.aspx Email James@amazon.com 2010/7/15 http://perspectives.mvdirona.com 19